Gemini 2.5: Updates to our family of thinking models

Source: DeepMind Date: 2025-06-17 URL: https://deepmind.google/blog/gemini-25-updates-to-our-family-of-thinking-models/

Summary

Google stabilized Gemini 2.5 Flash and introduced 2.5 Flash-Lite preview, while adjusting pricing: Flash input tokens doubled to $0.30/1M but output dropped to $2.50/1M, eliminating the thinking/non-thinking price split. Flash-Lite launches with thinking disabled by default for maximum latency efficiency. Adoption of 2.5 Pro confirmed at Cursor, Bolt, GitHub, and Replit.

Implications

The thinking-budget API is the real product. Adjustable thinking budgets — controllable by developers via API — is where Google is making a structural bet. It allows integrators to tune reasoning depth against cost, which is the right knob for production agent pipelines where most calls don’t need deep reasoning. This mirrors what Anthropic is doing with extended thinking on Claude.

Pricing signals competitive pressure. Doubling input token price while cutting output price is a deliberate move — it optimizes for agentic and long-context use cases (where output volume dominates) while slightly taxing simple retrieval calls. That’s an integrators play, not a consumer play.

Named integrations matter. Cursor, Bolt, GitHub, and Replit as named adopters of 2.5 Pro is the integrator distribution story Google needs. These are the same tools that are Claude’s primary distribution wedge. Every named integration there is a point where Google and Anthropic are directly competing for developer default.

Watch:

• Whether Flash-Lite’s quality holds against Gemini 1.5/2.0 Flash in real production evals by integrators
• Thinking budget API adoption — if integrators start exposing it as a config knob, that’s a new benchmark surface
• Claude’s response: Haiku 4.5 and Sonnet pricing in the same tier window